International Journal of Civil Engineering and Technology (IJCIET)
Volume 10, Issue 04, April 2019, pp. 1158–1170, Article ID: IJCIET_10_04_122
Available online at http://www.iaeme.com/ijmet/issues.asp?JType=IJCIET&VType=10&IType=4
ISSN Print: 0976-6308 and ISSN Online: 0976-6316
© IAEME Publication
Scopus Indexed
FUZZY KNOWLEDGE BASE SYNTHESIS OF
THE EXPERIENCE LEVEL CLASSIFICATION
OF AVIATION SECURITY SCREENERS USING
SUB-TRACTIVE CLUSTERING AND ANFISTRAINING
A. A. Gladkikh
Ministry of Education and Science of the Russian Federation, Ulyanovsk State Technical
University, Severnyj venec, 32, 432027, Ulyanovsk, Russia
An. K. Volkov, Al. K. Volkov, N.A. Andriyanov, L.V. Mironova
The Federal Air Transport Agency, Ulyanovsk Civil Aviation Institute, Mozhaisky Street,
8/8, 432071, Ulyanovsk, Russia
ABSTRACT
This paper describes a new approach to the estimate reliability improvement of the
competence related to the visual searching for prohibited items by aviation security
screeners consisting in the application of fuzzy knowledge bases and equipmentspecific diagnostic techniques of the psychophysiological state. As equipment-specific
methods it is suggested using the eye-tracking technology giving an opportunity to
take into consideration visual searching strategies and the variational cardiointervalometry method giving an opportunity to estimate the operator’s
psychophysiological strain. This paper considers the theoretical basis of the automatic
synthesis of both Sugeno and Mamdani fuzzy knowledge bases. It is demonstrated that
the identification of fuzzy knowledge bases using clustering algorithms consists of
cluster forming in the dataspace and cluster transforming into fuzzy rules describing
the certain part of the investigated system behavior. Experiments were carried out to
test the suggested approach. The quality of the synthetized Sugeno fuzzy knowledge
base was compared with the Mamdani knowledge base and with the linear regression
model. Results showed that this model based on the subtractive clustering and the
ANFIS-training with respect to the root-mean-square error does better apprise the
investigated dependence than other models. The training sample accuracy was
0.0049. The test sample accuracy was 0.0275. The application of fuzzy knowledge
bases will give an opportunity to create decision-making support systems to estimate
the competence level of aviation security screeners in intellectual simulator
complexes.
Key words: Aviation security, Aviation Security screener, Competence, Fuzzy
Models, Subtractive Clustering, Eye-Tracking Technology and Heart Rate.
http://www.iaeme.com/IJCIET/index.asp
1158
editor@iaeme.com
Fuzzy Knowledge Base Synthesis of the Experience Level Classification of Aviation Security
Screeners Using Sub-Tractive Clustering and Anfis-Training
Cite this Article: A.A. Gladkikh, An.K. Volkov, Al.K. Volkov, N.A. Andriyanov,
L.V. Mironova, Fuzzy Knowledge Base Synthesis of the Experience Level
Classification of Aviation Security Screeners Using Sub-Tractive Clustering and
Anfis-Training, International Journal of Civil Engineering and Technology 10(4),
2019, pp. 1158–1170.
http://www.iaeme.com/IJCIET/issues.asp?JType=IJCIET&VType=10&IType=4
1. INTRODUCTION
Despite of the constant improving of X-ray technologies and improving automation in
inspection systems, the human factor in the form of operators’ qualification and the
responsibility is an essential component of the efficient process structuring of inspecting in
airports. The primary goal of aviation security screeners is to decide whether there are any
dangerous and prohibited items and matters at the aircraft using X-ray images of the baggage
and carry-on. In connection therewith, today one of the most important and urgent tasks to be
resolved by the aviation industry is to improve the quality of the professional training of this
very category of the aviation personnel. The professional training system of aviation security
screeners is based on the simulator training technology. Today the simulator training of
operators is carried out using various simulators. X-ray Tutor (Halbherr et al. 2013; Michel et
al. 2014) [7, 12] is the most common. Implementation of psychophysiological state
monitoring mechanisms of trainees is one of promising trends in improving simulators for
aviation security screeners. It is suggested using the eye-tracking technology for the visual
searching strategy estimation and the variational cardio-intervalometry method for the
psychophysiological strain estimation (the psychophysiological “price” of the executing task)
as equipment-specific monitoring methods. Eye movement features investigation when
performing professional tasks and training are being actively developed in various functional
areas. Investigations has been significantly advanced when working with the oculomotor
activity of medical experts, in particular: surgeons when working together in one team (Atkins
et al. 2013; Causer et al. 2014) [2, 5]; radiologists when analyzing the X-ray image (Wood et
al. 2013) [21]; cardiologists when reading the ECG tracing (Wood et al. 2014; Bond et al.
2014) [20,3]. The series of investigations (Vrzakova et al. 2012; Weibel et al. 2012) [17, 18]
relate to using the eye-tracking technology for the estimation of vision perception parameters
and pilot’s attention allocation during the simulator training when resolving various tasks.
During the professional training of sportsmen, the eye-tracking technology can be used as an
instrument for resolving such tasks as molding of efficient strategies for the visual searching
or comparing oculomotor patterns of trainees with various levels of training (Hancock et al.
2013; Witte et al. 2012; Button et al. 2010; Vickers, 2007) [8, 19, 4, 16]. In investigations
(Swann et al. 2014; Swann et al. 2015) [14, 15] using the eye-tracking technology there is
activity structure of aviation security screeners identified, including: visual searching, image
evaluation, compatibility with the equipment interface, cooperation with other inspection
staff, compatibility with the investigated object. Thus, the implementation of equipmentspecific means for the psychophysiological monitoring will provide objective and
comprehensive information about the process of the operators’ simulator training. At the same
time, taking into consideration the increase in the amount of the processed information, it is
necessary to develop new approaches to support the decision-making in estimating the quality
of the operators’ simulator training of operators based on artificial intelligence systems. These
systems include the fuzzy logics apparatus, artificial neural networks, genetic algorithms.
Estimation of the results of the operators' simulator training often occurs under conditions of
the uncertainty and the incompleteness of the initial data, therefore, in this paper it is proposed
to apply fuzzy knowledge bases. In the work (Hayajneh et al. 2007) [10] the authors resolved
http://www.iaeme.com/IJCIET/index.asp
1159
editor@iaeme.com
A.A. Gladkikh, An.K. Volkov, Al.K. Volkov, N.A. Andriyanov, L.V. Mironova
the issue of identifying the nonlinear dependence in the well quality control task when drilling
based on the fuzzy knowledge base. In the work (Hayajneh et al. 2010) [9] the authors used
the fuzzy knowledge base as the basis of the system for monitoring tile defects when
producing. In the work (Priyono et al. 2012) [13] the result of fuzzy knowledge base
designing for the traffic manage-ment task is presented. The investigation (Aires et al. 2009)
[1] reflects the results of fuzzy knowledge bases designing for identifying malfunctions in the
oil refining industry.
The objective of this investigation is to substantiate the possibility of using fuzzy
knowledge bases and equipment-specific diagnostic techniques of the psychophysiological
state of operators during the simulator training.
1.1. Theoretical basis for designing fuzzy knowledge bases
Today, there are two main types of fuzzy knowledge bases: Mamdani and Sugeno. The
difference between Mamdani and Sugeno fuzzy inferences is since in the first case
conclusions of rules are represented by fuzzy terms as well as for input variables, and in the
second case conclusions are a function of input variables. Mamdani knowledge base is
described as follows (Mamdani and Assilian, 1975) [11]:
If ( x1  ai1 и x2  ai 2 и...и xn  ain ) , then y  d i , weighted wi , i  1, N ,
(1)
Where:
a ij is the fuzzy term used for the linguistic variable estimation of the factor x j in the irule, i  1, N , j  1, n ;
N is the number of rules in the knowledge base;
d i is the consequent of the i-rule in the form of the fuzzy term;
wi  0; 1 is the weight of the i-rule representing the expert’s certainty in its authenticity.
Zero-order Sugeno knowledge base is described as follows (Chiu, 1994) [6]:
If ( x1  ai1 и x2  ai 2 и...и xn  ain ) , then d j  b j 0   b ji xi ,
i 1, n
(2)
Where:
b j 0 , b j1 ,..., b jт are some real numbers.
Clustering methods, namely, the subtractive clustering algorithm or fuzzy c-means are
often used for the synthesis of fuzzy knowledge bases. Structuring a Sugeno fuzzy knowledge
base of experimental data consists of two main steps. The first is the identification of the
knowledge base structure, which includes the formation of the initial set of rules in “IF-TO”
statements using subtractive clustering. The second is the identification of parameters using
the ANFIS-algorithm (ANFIS is Adaptive Network Based Fuzzy Inference System), which
includes the setting of accession and weights of rules functional parameters. The use of
clustering gives an opportunity to identify natural data groups from the overall set, thereby
providing the identification and the brief presentation of the overall structure of the
knowledge base. The advantage of the subtractive clustering algorithm is the lack of need to
determine the initial number of clusters. The algorithm includes the following steps:
consideration of data elements as potential candidates for cluster centers; probability
calculation that the data element is the center of the cluster; selection of the point with the
greatest potential as the center of the first cluster; calculation of potentials for the next cluster
centers (deducting the contribution of the center of the cluster just found); implementation of
an iterative recalculation procedure for potentials and separation of cluster centers until the
maximum potential value exceeds the initially specified threshold (Yager and Filev, 1984)
http://www.iaeme.com/IJCIET/index.asp
1160
editor@iaeme.com
Fuzzy Knowledge Base Synthesis of the Experience Level Classification of Aviation Security
Screeners Using Sub-Tractive Clustering and Anfis-Training
[22]. Although the subtractive clustering algorithm is not fuzzy, it is often used for the task of
the automatic generation of fuzzy rules from experimental data.
Thus, the identification of fuzzy knowledge bases using subtractive clustering involves
forming of clusters in the data space and describing the certain part of the investigated system
behavior. After projecting the membership degrees of clusters on the input space and the
corresponding approximation, the membership functions of terms in the premises of fuzzy
rules are determined. Conclusions of rules are determined by the least square method.
At the second stage, parameters of the synthesized knowledge base are configured using
the ANFIS-algorithm. The representation of the fuzzy inference in the form of the neurofuzzy model gives an opportunity to the use neural network learning algorithms. The structure
of the ANFIS-network is isomorphic to the fuzzy knowledge base and is a five-layer, forwardpropagation neural network. Layers have the following assignments: the first layer is the term
of input variables; the second layer is sending fuzzy rules; the third layer is the normalization
of the compliance degree of rules; the fourth layer is the conclusion of rules; the fifth layer is
the aggregation of the results obtained by various rules. Combination of the back-propagation
of error algorithm and the least square method are often used as training procedures for these
models.
The fuzzy c-means algorithm is used to design the Mamdani fuzzy knowledge base use.
Clustering in this case can be formulated using the characteristic function. The characteristic
function takes the value in the interval [0, 1] reflecting the degree of belonging of the element
to the cluster. The description of the partitioning matrix of fuzzy clusters using a characteristic
function is represented as F  [ ki ] , where ki  [0,1], k  1, M , i  1, c . In this case, the k-line
characterizes the degree of belonging of the element X k  ( xk1 , xk 2 ,..., xkn ) to clusters
A1 , A2 ,..., Ac . The F-matrix must satisfy the following requirements (Mamdani and Assilian,
1975) [11]:
  ki  1, k  1, M ,
i 1, c
(3)
0    ki  M , i  1, c.
k 1, M
The idea of the fuzzy c-means algorithm is to iteratively recalculate the F-matrices and
cluster centers. The scatter criterion is used as the target function, which must be minimized
(Mamdani and Assilian, 1975) [11]:
m
 (  ki ) Vi  X k

2
 min,
i 1, c k 1, M
(4)
Where:
m
 ( ki ) X k
Vi 
k 1, M
m
 ( ki )
are fuzzy cluster centers; m  (1, ) is the exponential weight.
k 1, M
In this case, the Euclidean distance is taken as the metric of the anomaly. Membership
functions of output variable terms in conclusions of rules of the knowledge base are found as
well as for input variables.
2. TASK SUBSTITUTION
Synthesizing the fuzzy knowledge base for resolving the issue related to the estimation of the
competence level of aviation security screeners is to identify fuzzy rules relating input data (x)
to the output (y):
http://www.iaeme.com/IJCIET/index.asp
1161
editor@iaeme.com
A.A. Gladkikh, An.K. Volkov, Al.K. Volkov, N.A. Andriyanov, L.V. Mironova
( X r , y r ), r  1, M ,
(5)
Where:
X r is the input data vector related to r-line of the sample and y r is the value of the output
variable.
Setting up the fuzzy knowledge base is carried out by the criterion of the root-meansquare error (RMSE):
RMSE 
1
M
 ( y r  F ( P,W , X r )) ,
(6)
r 1. M
Where:
M is the number of experimental data pairs; P is the parameter vector of variable
membership functions (x) and (y); W is the knowledge base weight coefficient vector;
F ( P,W , X r ) is the result of the fuzzy knowledge base calculation.
As input data (x), it is proposed to use (3) the integral indicators characterizing the
strategies of visual searching for prohibited items, and (4) the indicator characterizing the
psychophysiological tension of the operator during the visual searching based on the heart rate
estimation.
1. The first indicator (DT) is calculated using the formula (7):
n
i 1
1
FC pro
1
n
j 1
k
DT  
 DTpro 
m
1
  DTpro 
 DTnonpro 
FC nonpro j 1
FC pro
l 1
,
(7)
Where:
DTpro is the average vision holding time of the test person in AOI (area of interest) with
the prohibited item, ms;
DTnonpro is the average vision holding time of the test person in AOI without the prohibited
item, ms;
1/FCpro is monitoring frequency of the j-prohibited item;
1/FCnonpro is monitoring frequency of the l-nonprohibited item;
k is the number of test X-ray images, pcs;
n is the number of AOI with the prohibited item, pcs;
m is the number of AOI without the prohibited item, pcs.
2. The second indicator (SE) is calculated using the formula (8):
k
SE 
 SEpro
i 1
k
,
(8)
Where:
SEpro is the index characterizing the concentration fixation scheme of the test person equal
to 0 or 1. If the test person began the searching scheme from AOI with the dangerous and
prohibited item or matter, the indicator will be equal to 1. Otherwise it will be equal to 0;
k is the number of test X-ray images, pcs.
3. The third indicator (ET) is calculated using the formula (9):
http://www.iaeme.com/IJCIET/index.asp
1162
editor@iaeme.com
Fuzzy Knowledge Base Synthesis of the Experience Level Classification of Aviation Security
Screeners Using Sub-Tractive Clustering and Anfis-Training
n


 ETpro  S pro


k
j 1

,
ET   m
n

i 1
  ETnonpro  S nonpro   ETpro  S pro 
j 1
 l 1

(9)
Where:
ETpro is the average time from the beginning of the experiment to the first fixation in AOI
with the prohibited item, ms;
ETnonpro is the average time from the beginning of the experiment to the first fixation in
AOI without the prohibited item, ms;
Spro is the part of the j-dangerous item area against the total area of the stimulus;
Snonpro is the part of the l-not dangerous item area against the total area of the stimulus;
k is the number of test X-ray images, pcs;
n is the number of AOI with the prohibited item, pcs;
m is the number of AOI without the prohibited item, pcs.
4. The indicator (HR) is calculated as follows:
HR 
Cur
YHR
BG
YHR
,
(10)
Where:
Cur
BG
Y HR
is the current value of the operator’s heart rate and YHR
is the background value of
the heart rate.
Detection rate of prohibited items is used as the output variable (DP).
To test the proposed approach, experimental investigations were carried out using the
mobile eye-tracker Eye Tracking Glasses 2.0 of Senso Motor Instruments. To assess the heart
rate, a set of psychophysiological testing devices UPFT-1/30 “Psychophysiologist” was used.
35 cadets of the Ulyanovsk Civil Aviation Institute took part in these investigations. They
were trained at the Aviation Training Center of the Institute under the 40-hour program
“Retraining of cadets in the department of aviation security. Preflight and post-flight
inspection”. The collection of experimental data included the following steps: block forming
of test X-ray images of baggage and carry-on consisting of 20 images; preprocessing of test
images using SMI BeGaze 3.7 software; experimental testing of test persons with the
simultaneous removal of the oculomotor activity, heart rate and fixing the time and accuracy
of the object identification (images were presented with the exposure of 10 seconds);
processing and analyzing of the experimental data. The initial sample was divided into the
training (30 cadets) and the testing sample (5 cadets).
3. RESULTS AND DISCUSSION
The robust estimation of the initial data matrix (their suitability for the subsequent analysis
and the absence of outliers) was carried out using the Statistica program. Table 1 presents a
window of descriptive statistics of the experimental data obtained.
http://www.iaeme.com/IJCIET/index.asp
1163
editor@iaeme.com
A.A. Gladkikh, An.K. Volkov, Al.K. Volkov, N.A. Andriyanov, L.V. Mironova
Table 1 Results of the robust estimation of initial data
In
dex
Mean
Trim
med
mean,
5,000
%
DT
SE
2.981
0.106
2.989
0.100
Win
sorized
mean,
5,000
%
2.983
0.106
ET
2.286
2.258
HR
0.953
0.958
Grubbs
Test
Statistic
pvalue
Minimum
Maximum
Std.
Dev.
2.208
2.006
0.795
1.000
1.999
0.000
3.781
0.300
0.445
0.097
2.294
2.603
0.220
0.623
4.096
0.695
0.955
3.037
0.039
0.805
1.000
0.049
Std.
Er
ror,
%
7.524
1.637
11.75
3
0.823
The robust estimation in the Statistica program was carried out according to three
indicators (Swann et al. 2014) [14]: the truncated mean (the average value after the removal
of outliers); the winsorized mean (the average value after replacing outliers with the
percentile, which made the truncation); the Grubbs criterion for outliers.
It can be seen from Table 1 that the arithmetic mean, the truncated mean and the
winsorized mean have approximately equal values for each of the estimated parameters. The
critical value of the Grubbs criterion for a significance level of 0.01 is 3.33. For the analyzed
parameters, Grubbs criteria are 2.207525; 2.006237; 2.603334; 3.036886; 2.001045 and do
not exceed the critical value; levels of significance equal to 0.795209; 1; 0.219935; 0.038577;
1 respectively exceed the selected level. Thus, the analysis performed allowed us to
substantiate the absence of outliers in the initial data and their suitability for the subsequent
analysis.
The factor analysis of the initial data matrix using the method of isolating the main
components shows that all indicators used bear enough factor load and can be used to
synthesize the fuzzy knowledge base (Table 2).
Table 2 Matrix of factor loads (the «varimax» method)
Index
DT
SE
ET
HR
Expl. Var.
Prp. Totl.
Factor 1
0.937
0.160
-0.033
-0.337
1.021
0.255
Factor 2
0.038
0.182
-0.953
0.281
1.022
0.256
Factor 3
0.182
0.967
-0.191
0.091
1.014
0.253
Factor 4
0.290
-0.072
0.232
-0.894
0.943
0.236
The Matlab Fuzzy Logic Toolbox package was chosen as the software environment for
implementing the fuzzy knowledge base. The procedure for synthesizing the Sugeno fuzzy
knowledge base is implemented in this package as the genfis2 function. Set the following
parameters of the subtractive clustering algorithm: the cluster radius is 0.7 (the value is set
from the range [0, 1]); the suppression ratio is 1.25; the adoption rate is 0.5; the rejection
coefficient is 0.15. Figure 1 shows the scheme of the synthesized Sugeno fuzzy knowledge
base.
http://www.iaeme.com/IJCIET/index.asp
1164
editor@iaeme.com
Fuzzy Knowledge Base Synthesis of the Experience Level Classification of Aviation Security
Screeners Using Sub-Tractive Clustering and Anfis-Training
Figure 1. Scheme of the synthesized Sugeno fuzzy knowledge base
The resulting knowledge base includes 3 rules that correspond to three found clusters:
IF DT=in1cluster1 AND SE=in2cluster1 AND ET=in3cluster1 AND HR=in4cluster1, THEN
DP=out1cluster1;
IF DT=in1cluster2 AND SE=in2cluster2 AND ET=in3cluster2 AND HR=in4cluster2, THEN
DP=out1cluster2;
IF DT=in1cluster3 AND SE=in2cluster3 AND ET=in3cluster3 AND HR=in4cluster3, THEN
DP=out1cluster3.
The error on the training sample is equal to trnRMSE1 = 0.0084. The error on the test
sample (outside training points) is equal to chkRMSE1 = 0.0445. Figure 2a presents the
results of comparing the value of the output variable from the experimental data and the
simulation results.
a)
b)
Figure 2. Testing results of the model
a) after the subtractive clustering; b) after the ANFIS-training
In some cases, there are discrepancies between the experimental data and the simulation
results (Figure 2a). To improve the accuracy of the model, we will train it using the ANFISalgorithm. Let us assign the value of learning iterations equal to 18. Based on the graph in
Figure 2b, we can conclude that the quality of the simulation has increased. Figure 3 shows
the dependence of modeling errors on the number of iterations of the ANFIS-algorithm.
http://www.iaeme.com/IJCIET/index.asp
1165
editor@iaeme.com
A.A. Gladkikh, An.K. Volkov, Al.K. Volkov, N.A. Andriyanov, L.V. Mironova
Figure 3. Dependence of modeling errors on the number of iterations of the ANFIS-algorithm
The errors after the ANFIS-training were equal to trnRMSE2 = 0.0049 and chkRMSE2 =
0.0275. Thus, even a small training allowed to increase the accuracy of the model. Analysis of
the training dynamics (Figure 3) allows us to conclude that the error on the test sample
reaches the lowest value at the 7th iteration (chkRMSE = 0.0246). At the same time, the error
on the training sample decreases throughout all 18 iterations.
The membership functions in the hypothesis of rules are described by the following
Gaussian function [17]:
μ( x)  e

( x  b) 2
2c 2
(11)
,
Where:
b is the coordinate of the membership function maximum and c is the coefficient of the
membership function concentration.
Example of the form of the membership function for the input variable HR of fuzzy
clusters before the training by the ANFIS-algorithm and after it is shown in Figure 4.
a)
b)
Figure 4. Membership functions for the input variable HR of fuzzy clusters
a) before training; b) after the ANFIS-training
Parameters of all membership functions are presented in Table 3.
http://www.iaeme.com/IJCIET/index.asp
1166
editor@iaeme.com
Fuzzy Knowledge Base Synthesis of the Experience Level Classification of Aviation Security
Screeners Using Sub-Tractive Clustering and Anfis-Training
Table 3 Parameters of membership functions of the Sugeno fuzzy knowledge base
Input
variable
DT
SE
ET
HR
Cluster
cluster1
cluster2
cluster3
cluster1
cluster2
cluster3
cluster1
cluster2
cluster3
cluster1
cluster2
cluster3
Before training
After the ANFIS-training
b
c
b’
c’
2.668
3.413
3.573
0.1
0.1
0
1.967
1.865
3.452
0.987
0.956
0.865
0.441
0.441
0.441
0.074
0.074
0.074
0.859
0.859
0.859
0.035
0.035
0.035
2.671
3.413
3.572
0.077
0.126
0.0005
1.967
1.86
3.452
0.986
0.945
0.866
0.443
0.444
0.441
0.066
0.070
0.075
0.861
0.857
0.860
0.039
0.023
0.038
Reported values of conclusions of rights for the Sugeno knowledge base are presented in
Table 4.
Table 4 Conclusions of the Sugeno knowledge base
Output variable (DP)
Cluster
Before training
= 0.31+0.9432·DT–0.2238·SE
–0.01595·ET+0.4891·HR
After the ANFIS-training
= 1.281+0.02203·DT–0.01662·SE
–0.00039·ET–0.3849·HR
cluster2
= –0.005155+0.144·DT+0.1199·SE+
+0.001565·ET+0.3977·HR
= –0.154+0.1672·DT+0.1938·SE+
+0.003474·ET+0.4531·HR
cluster3
= 0.7259+0.003549·DT–2.492·SE+
+0.0007605·ET+0.126·HR
= 0.7096+0.00613·DT+3.61·SE+
+0.002314·ET+0.1281·HR
cluster1
Based on Table 3 and Table 4, it can be seen that because of the ANFIS-training, both
parameters of the membership functions and parameters in conclusions of fuzzy inference
rules were refined.
Let us compare the quality of the synthesized Sugeno model with the Mamdani fuzzy
knowledge base and the linear regression. Synthesis of the Mamdani knowledge base is
accomplished using the genfis3 function. Let us assign the following parameters of the fuzzy
c-means algorithm: the number of clusters is 3; the exponential weight is 2; the value of the
improvement of the objective function for one iteration is 0.00001; the number of iterations is
100. As a result, the fuzzy knowledge base, which also contains 3 fuzzy rules, is extracted
from the experimental data. Error values are equal to trnRMSE3 = 0.0304 and chkRMSE3 =
0.0349.
The quality of the identified linear regression model according to the adjusted
determination coefficient of was R 2  0.83 . Coefficients of this model and its characteristics
are presented in Table 5.
Table 5 Regression model
http://www.iaeme.com/IJCIET/index.asp
1167
editor@iaeme.com
A.A. Gladkikh, An.K. Volkov, Al.K. Volkov, N.A. Andriyanov, L.V. Mironova
Variable
Coefficient
Standard error
Constant
DT
SE
ET
HR
0.554
-0.025
0.095
-0.005
0.467
0.103
0.007
0.034
0.005
0.084
Value of Student
t-criterion
5.332
-2.874
2.873
-0.920
5.595
Test significance
(p-value)
9.102E-06
0.007
0.008
0.364
4,339E-06
The model error values were trnRMSE4 = 0.0371 and chkRMSE4 = 0.0313.
To assess the splitting quality of the source data into 3 fuzzy clusters, we calculate the
Xei-Beni index using the following formula (Yager and Filev, 1984) [22]:

χ
m
 ( ki ) X k  Vi
i 1, c k 1, M
2
M min ( X k  Vi )
2
.
(12)
i j
The calculated value of the Xei-Beni index by the formula 8 was 0.5348. The optimal
splitting into clusters meets the Xei-Beni criterion less than 1. Thus, we can conclude that a
good result has been obtained for splitting into fuzzy clusters.
From Table 6 it can be seen that the Sugeno fuzzy knowledge base based on the
subtractive clustering and the ANFIS-training do better approximate the dependence between
the indicators of the operator's visual searching strategies, the indicator of
psychophysiological tension and the detection efficiency of prohibited items than other
models.
Table 6 Comparison results of obtained models
Model
RMSE on training sample
RMSE on test sample
Sugeno fuzzy model (without the ANFIStraining)
0.0084
0.0445
Sugeno fuzzy model (with the ANFIStraining)
Mamdani fuzzy model
0.0049
0.0275
0.0304
0.0349
Linear regression model
0.0371
0.0313
4. CONCLUSIONS
This paper proposes a new approach to estimating the competence level of aviation security
screeners allowing to take into consideration parameters of the oculomotor activity and the
heart rate variability of the test operators, which differs from the existing approaches using
fuzzy classification models. According to the results of an experimental investigation, three
different models were synthesized. Comparison results showed that the Sugeno model based
on the subtractive clustering and the ANFIS-training is more accurate than the Mamdani
model and regression models.
Monitoring mechanism integration of the psychophysiological state of operators into the
process of the simulator training will give an opportunity to form the comprehensive
conclusion about the degree of development of their professional competencies and readiness
for real practical activities. The results of this monitoring can be used to improve the selection
of the operator's training strategy based on not only training indicators, but also using the socalled psychophysiological "price" of the test person activity. The introduction of the
additional feedback loop into adaptive training algorithms based on the use of eye-tracking
technology will give an opportunity to instructors to track and to timely adjust the activity of
http://www.iaeme.com/IJCIET/index.asp
1168
editor@iaeme.com
Fuzzy Knowledge Base Synthesis of the Experience Level Classification of Aviation Security
Screeners Using Sub-Tractive Clustering and Anfis-Training
the trained operator in real time. Application of fuzzy knowledge bases will give an
opportunity to create decision-making sup-port systems for the estimation of the competence
level of aviation security screeners in intelligent training complexes.
In addition to the noted, the proposed approach can be used to resolve the task of
determining the risk level in the airport inspection system. In this case, physical and
technological parameters of inspected objects and relevant characteristics of equipmentspecific inspection methods can be considered as input parameters of the fuzzy logical
inference. For example, physical indicators of the prohibited item can be density, mass,
dimensions, volatility, etc. Characteristics of equipment-specific inspection methods can
include detection probability, false alarm probability, throughput, reliability, etc. At the same
time, this system is considered as an ergatic one. That is why the key role is assigned to the
human factor, which is characterized by the reliability of the human operator. Probability of
the fact that prohibited items and substances pass the inspection, can be considered as an
integral risk indicator. Application of fuzzy knowledge bases will increase the interpretability
of the risk model and identify those indicators which value has mostly affected the risk level.
Eventually, this will reduce costs of the aviation security audit.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
Aires, L., Araujo, J., Dourado, A. Industrial monitoring by evolving fuzzy systems. Joint
World Congress of the International Fuzzy Systems Association and Conference of the
European Society for Fuzzy Logic and Technology. Lisbon. 2009. pp. 1358-1363.
Atkins, M., Tien, G., Khan, R., Meneghetti, A., Zheng, B. What do surgeons see:
capturing and synchronizing eye gaze for surgery applications. Surgical Innovation, 20(3),
2013, pp. 241–248.
Bond, R. R., Zhu, T., Finlay, D. D., Drew, B., Kligfield, P. D., Duldenring, D., Breen, C.
Gallagher, A. G., Daly, M. J., Clifford, G. D. Assessing computerized eye tracking
technology for gaining insight into expert interpretation of the 12-lead electrocardiogram:
an objective quantitative approach. Electrocardiology, 47(6), 2014. pp. 895–906.
Button, C., Dicks, M., Haines, R., Barker, R., Davids, K. Statistical modeling of gaze
behavior as categorical time series: what you should watch to save soccer penalties.
Cognitive Processing. 12(3). 2010. pp. 235-244.
Causer, J., Vickers, J.N., Snelgrove, R., Arsenault, G., Harvey, A. Performing under
pressure: quiet eye training improves surgical knot-tying performance. Surgery, 156(5),
2014, pp. 1089–1096.
Chiu, S. L. Fuzzy Model Identification Based on Cluster Estimation. Journal of Intelligent
and Fuzzy Systems. 2. 1994. pp. 267– 278.
Halbherr, T., Schwaninger, A., Budgell, G., Wales, A. Airport security screener
competency: a cross-sectional and longitudinal analysis. International Journal of Aviation
Psychology, 23(2), 2013, pp. 113-129.
Hancock, D.J., Ste-Marie, D.M. Gaze behaviors and decision making accuracy of higherand lower-level referees. Psychology of Sport and Exercise, 2013, 14, pp. 66-71.
Hayajneh, M., Hassan, A. Monitoring defects of ceramic tiles using fuzzy subtractive
clustering-based system identification method. Soft Computing.
14 (6). 2010. pp. 615626.
Hayajneh, M., Radaideh, S. Monitoring Hole Quality in a Drilling Process Using a Fuzzy
Subtractive Clutering-based System Identification Method. Journal of Testing and
Evaluation. 35 (3). 2007. pp. 287-302.
Mamdani, E. H., Assilian, S. An Experiment in Linguistic Synthesis with Fuzzy Logic
Controller. International Journal Man-Machine Studies. 7 (1). 1975. pp .1–13.
http://www.iaeme.com/IJCIET/index.asp
1169
editor@iaeme.com
A.A. Gladkikh, An.K. Volkov, Al.K. Volkov, N.A. Andriyanov, L.V. Mironova
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
Michel, S., Mendes, M., de Ruiter, J., Koomen, G., Schwaninger, A. Increasing X-ray
image interpretation competency of cargo security screeners. International Journal of
Industrial Ergonomics, 27, 2014, pp. 551-560.
Priyono, A., Ridwan, M., Alias, A. J., Rahmat, R. A. O. K., Hassan, A., Ali, M. A. M.
Generation of fuzzy rules with subtractive clustering. Journal Teknologi. 43 (1). 2012. pp.
143-153.
Swann, L., Popovic, V., Blackler, A. L., Kraal, B. J. Airport security screeners expertise
and implications for interface design. Design Research Society Conference, Umea, 2014.
Swann, L., Popovic, V., Thompson, H., Blacker, A., Kraal., B. Relationships between user
experience and intuitiveness of visual and physical interactions. Proceedings of the 6th The
International Association of Societies of Design Research. Brisbane, 2015 pp. 1900-1916.
Vickers, J. N. Mind over muscle: The role of gaze control, spatial cognition, and the quiet
eye in the motor expertise. Cognitive Processing. 12. 2007. pp. 219-222.
Vrzakova, H., Bednarik, R. Hard lessons learned: mobile eye-tracking in cockpits.
Proceedings of 4-th Workshop on eye gaze in intelligent human machine interaction,
ACM, 7, 2012, pp. 130-136.
Weibel, N., Fouse, A., Emmenegger, C., Kimmich, S., Hutchins, E. Let’s look at the
cockpit: exploring mobile eye-tracking for observational research on the flight deck.
Proceedings of the Symposium on Eye Tracking Research and Applications, Warsaw,
2012, pp. 107-114.
Witte, K., Emmermacher, P., Bandow., N., Masik, S. Usage of Virtual Reality Technology
to Study Reactions in Karate-Kumute. International Journal of Sports Science and
Engineering. 6(1). 2012. pp. 17-24.
Wood, G., Batt, J., Appelboam, A., Harris, A., Wilson, M. R. Exploring the impant of
expertise, clinical history, and visual search on electrocardiogram interpretation. Medical
Decision Making, 34(1), 2014, pp. 75–83.
Wood, G., Knapp, K.M., Rock, B., Cousens, C., Roobotton, C., Wilson, M.R. Visual
expertise in detecting and diagnosing skeletal fractures. Skeletal Radiology, 42(2), 2013,
pp. 165–172.
Yager, R., Filev, D. Essentials of Fuzzy Modeling and Control. New York : John Wiley &
Sons, 1984, pp. 387.
http://www.iaeme.com/IJCIET/index.asp
1170
editor@iaeme.com